Maximum a posteriori estimation of diffusion tensor parameters using a Rician noise model: why, how and but

Area of Science:

• Medical Imaging
• Biophysics
• Computational Neuroscience

Background:

• Diffusion-weighted MRI (DW-MRI) commonly uses the diffusion tensor model.
• Standard parameter estimation assumes Gaussian or log-normal noise, inadequate for high b-values or low SNR.
• This leads to biased diffusion parameter estimates, underestimating true diffusion.

Purpose of the Study:

• To present a computational framework for estimating diffusion tensor parameters using a Rician noise model.
• To address biases in diffusion parameter estimation caused by inadequate noise models in DW-MRI.
• To provide a robust method for quantitative analysis in challenging imaging conditions.

Main Methods:

• Developed a computational framework utilizing a Rician noise model for diffusion tensor spectral decomposition.
• Employed a Fisher-scoring scheme for robust and rapid parameter estimation.
• Investigated the utility of the Fisher-information matrix for optimal experimental design.

Main Results:

• The Rician noise model yields significantly less biased diffusion parameter estimates across various b-values and SNR levels.
• Rician estimates show poorer precision than Gaussian models at very low SNR.
• Pooling Rician uncertainty estimates improves precision, though still less than Gaussian models.

Conclusions:

• The Rician estimator is recommended for applications requiring truly quantitative diffusion MRI values and predictive model comparisons.
• Gaussian estimators may be preferable for longitudinal or group comparisons where relative precision is paramount.
• The proposed framework enhances the reliability of diffusion tensor imaging analysis, particularly under suboptimal imaging conditions.